AI and energy infrastructure, explained plainly.
A journal on the physical constraints behind the AI buildout: power, grid connection, data centres, cooling and the supply chains that gate them. Written to be understood by anyone.
Almost every watt that goes into an AI chip comes out as heat, and one cabinet now gives off as much as eighty space heaters. Air can't carry that away. So the biggest plumbing change in the history of data centres is happening right now, and it doesn't care which chip company wins.
When an AI writes its reply, the arithmetic is fast and the waiting for data is slow. So the industry started stacking memory chips into little towers glued right next to the processor. It's a clever fix, it's expensive, and it explains a lot about which parts of the AI supply chain are actually scarce.
A frontier AI model is too big to fit on any single chip, so it's sliced across thousands of them, and they have to swap notes for every single word. If the wiring between them is even slightly slow, the most expensive chips ever built sit idle. That's why a huge slice of every AI dollar goes on cables and light.
AI computers are getting so power-hungry that the copper bars carrying electricity inside each cabinet would soon weigh more than the computers. So the industry is about to change how power enters the building. That change, not the chips, is what decides who gets to build.
When you ask an AI a question, no answer is waiting on a shelf. A building the size of a factory manufactures every word for you, one at a time, in real time. Understanding that one fact explains why the whole industry is suddenly about electricity, water and copper.
Everyone watches the chip supply. Almost nobody watches the queue to plug a data centre into the grid. But a chip order arrives in months, and a substation takes years. The scarce thing is a finished connection to the wires, and it isn't priced like it's scarce.
